Meta Voicebox

What is Meta Voicebox?

Meta Voicebox is a speech generation model developed by Meta. It is built on a non-autoregressive flow-matching architecture and is trained to fill in missing speech using both existing audio and provided text. Because it learns from context, Voicebox outperforms AI models designed for a single specific speech task.

Voicebox can generate speech in six languages, remove background noise, edit audio content, transfer speech styles within and across languages, and generate varied speech samples quickly — up to 20 times faster than older, autoregressive (step-by-step) models.

Who created Meta Voicebox?

Voicebox was created by Meta. It is a speech generative model built on Meta’s non-autoregressive flow matching architecture, capable of text-to-speech, noise removal, content editing, style transfer, and varied sample generation across six languages. The available information does not name specific individuals behind the project.

Who is Meta Voicebox for?

• Content creators (videos, social media)
• Audiobook narrators
• Podcasters
• Voice actors
• Language teachers and tutors
• Customer service teams
• Game developers
• Film and animation professionals
• Speech therapists
• Translators

How to use Meta Voicebox?

1. Understand the model: Voicebox is a non-autoregressive flow-matching model trained to fill in speech gaps using audio context and text. Unlike step-by-step models, it can use information from both before and after the segment you’re working on. It supports text-to-speech within a language or across languages without task-specific training data (zero-shot), plus style transfer, noise removal, content editing, and varied sample generation.
2. Explore the demos: The Voicebox website hosts examples of its editing, sampling, and style transfer features, including cross-language capabilities. Browsing these demos gives a clear sense of what the tool can do.
3. Remove unwanted noise: Voicebox can remove sudden noises from recordings — for example, a doorbell or a barking dog — so you don’t have to re-record, helping recordings sound cleaner and more professional.

Meta Voicebox is a versatile speech generation model for generating and manipulating multilingual speech guided by text.